Multi-Scales Hybrid Graph Attention Neural Network with Jaccard Correlation Embedding for Link Prediction

By predicting the existence of a link in networks, link prediction finds widespread applications in power grid dispatching and security. Recently, Graph Auto-Encoders (GAE) based link prediction methods effectively integrate deep learning into link prediction tasks and attract much attention. However, the existence of a link between two nodes may be closely related to their neighboring nodes in different scales, when GAE-based encoding processes only extract features from the adjacency matrix at a single scale, establishing node features that account for varying influences from different neighborhood sizes becomes challenging. To alleviate this situation, this paper proposes a Multi-scales Hybrid Graph Attention Neural Network with Jaccard Correlation Embedding (MHGAT) to address link prediction tasks. Specifically, MHGAT enhances graph node features by integrating Jaccard Correlation, which reflects the neighboring information, as heuristic information. Then, it designs a multi-scale hybrid attention method for extracting link features around multiple neighborhoods for building a graph encoder to model the importance of nodes. By incorporating a designed node correlation-based graph auto-encoding loss, MHGAT extracts multi-scale neighboring correlation features and leverages these features for link prediction. Experimental results demonstrate that compared to commonly used link prediction methods, the proposed MHGAT approach achieves superior performance. The experiments also validate the effectiveness of the designed multi-scale hybrid attention and node correlation embedding.